AI, ML & Neural Networks — explained without jargon

The broadest field — building systems that perform tasks we'd normally call 'intelligent': perception, reasoning, planning, language, and decision-making.

• Umbrella term — includes rule-based expert systems, search algorithms, robotics, and modern ML.
• Goal: make machines act rationally given inputs and goals, not necessarily 'think like humans'.
• Two loose camps: Narrow AI (today's models, great at one task) and AGI (hypothetical, general-purpose).
• Every LLM, chatbot, recommender, and self-driving stack is a kind of AI system.

A subset of AI where systems learn patterns from data instead of being explicitly programmed with rules.

• Supervised learning — learn a mapping from labeled examples (e.g., spam vs. not-spam).
• Unsupervised learning — find structure in unlabeled data (clustering, embeddings).
• Reinforcement learning — learn by trial-and-error to maximize a reward signal.
• Workflow: data → features → model → loss function → optimizer → evaluation.

A family of ML models loosely inspired by the brain — stacked layers of simple units (neurons) that transform inputs into useful representations.

• Each neuron computes a weighted sum of inputs, adds a bias, and applies a non-linear activation.
• Layers compose: early layers learn low-level features, deeper layers learn abstract concepts.
• Trained via backpropagation + gradient descent on a loss function.
• Deep learning = neural networks with many layers; Transformers (the backbone of LLMs) are a specific NN architecture.